Why InceptionResnetV2 has 1536 output feature maps

Cadene / pretrained-models.pytorch

Pretrained ConvNets for pytorch: NASNet, ResNeXt, ResNet, InceptionV4, InceptionResnetV2, Xception, DPN, etc.

BSD 3-Clause "New" or "Revised" License

9.03k stars 1.83k forks source link

Why InceptionResnetV2 has 1536 output feature maps #206

Open Podidiving opened 3 years ago

Podidiving commented 3 years ago

https://github.com/Cadene/pretrained-models.pytorch/blob/8aae3d8f1135b6b13fed79c1d431e3449fdbf6e0/pretrainedmodels/models/inceptionresnetv2.py#L300

In this realization model has 1536 feature maps in the end.

However, according to https://arxiv.org/pdf/1602.07261.pdf it should be 2048 (fig 19).

Please, can you explain, why the realization is slightly different from the one, provided in the paper?

Rupesh-rkgit commented 2 years ago

I tried to understand the model and came to the conclusion that the whole model has followed the "schema of the Inception-v4 network", not InceptionResnetV2. Just see fig 9. It has also followed Inception A, Inception B, and Inception C block structure as we can see in fig 4,5,6.